Datensatz

Zusammenfassung

Seismological phase detection is a fundamental part of seismic observation process. The wave arrival tracking is complicated in sparse seismic arrays over low-seismicity areas, where the noise floor is high and signal-to-noise ratios are low. Classic automatic routines can fail to detect or falsely detect too many of the small induced seismic events. In this research we compare the results of multiple automatic and machine-learning based phase picking algorithms around Baltic States. In a region of low seismicity, we intend to track induced events, occurring from military exercises and mining blasts. We analyze 11 seismological station data of year 2021 and compare manual phase picks against automatically detected ones. For phase detection we use variations of STA/LTA and deep neural network pickers from Ross et al., 2018 and Zhu & Beroza, 2018. We show that automatic picking routines coincide very well with manual observations. Even regionally untrained neural networks outperform manual observations, resulting in increased event detections. Neural networks picks are also less susceptible to seismic noise than classic STA/LTA variations, however, false-detection problem cannot be eliminated entirely. We believe that improvements to neural network models could be made, retraining the model using local data or adjusting hyperparameters to better fit the low signal-to-noise ratios of the induced events.